DFE interpretation questions?

[completelyuseless]

hi all, im not very bright and ive been struggling with this for some time.
the DFE that im using has a feedforward filter and a feedback FIR filter.

1) my received signal (from a rayleigh fading channel) is input into the feedforward FIR filter. the output of the received signal is y(t)

2) i then subtract the output of my feedback filter from y(t) - x(t) = d(t)

3) d(t) goes into a decision device ==> d'(t) and then into the feeback filter, from which i get x(t) in step 2.

which signal (or combination thereof) would represent the equalization error?
which signal is the ideal signal?

any help would be greatly appreciated.

 

[bulx]

> hi all, im not very bright and ive been struggling with...
It is fairly obvious that real learning is a result of the struggling process..dont lose heart

Guess you need to compare to the original transmitted signal to find any sort of error. So a measure of error would E [(d'(t) - s(t))^2], where s(t) is the transmitted symbol (before the ralyleigh channel), which itself is the "ideal signal".

-b

 

[completelyuseless]

ok, this ive seen before (or something similar atleast) except that to get the equalization error, they use d(t), and not d'(t)... it doesnt matter, ive tried both. however, whats confusing me is how does the receiver know what the transmitted signal is before the rayleigh channel?


is there and estimation done on the received signal that provides the "ideal signal" in real world applications?
is the purpose of the feedback filter to create an estimation?

 

[bulx]

In reality, the reciever cannot estimate the actual error, except in circumstances where the transmit symbols are known beforehand. (It is quite common to know the trasmistted symbols during the setup phase of a comm system). Moreover, if it always knew the exact error, it could have got back the sent data without any error always! Practically, it will re-train whenever known symbols are sent, and assumes that the estimates are good enough to be applied when unknown data is being sent.

Also, it is useful to assume that the sent signal is known at the Rx and thus arrive at an error estimate by analysis or simulation; and use it as a criterion to compare different equalisation methods or to find ways of improving.

The feedback part is trying to remove the effect of past symbols on the current symbol by subtracting an optimally weighted sum of few past samples.

-b

 

[completelyuseless]

okay, that makes sence, the only thing thats bothering me now is this.

the question i should really ask is what should i see happening in the tap weight vector.

a while ago i did a simulation of an adaptive filter. (a plain and simple FIR feed forward filter.) the input signal was a sine wave with noise, and hence the desired signal was a pure sine wave. after the coefficients adapted, they remain the same, i thought that this was due to the single being a periodic signal. so if the ideal signal is removed, and the coefficients are unchanged, then the system will still work.

in my DFE, im using randomly generated bits that will form symbols for DQPSK. because its random, whenever i look at the tap coefficient vector, its always changing. i assume this is happening because the signal isnt periodic. but so far, im only getting a relativly correct output becuase i have the known signal.

so what exactly is used when the desired signal is unknown?

 

[cheggy]

Hi,

Tap coefficients change because of residual error (I suppose that your equaliser has converged, and these changes are just "oscillations" around some tap values). Have you checked results of equalisation: is data leaving equaliser (after decision is made) good or not?

 

[completelyuseless]

hi, sorry for such a late reply, and i really appreciate all your input.


there are two equalization algorithms that i am using,
RLS, and
RLS with adaptive forgetting factor.

the plain RLS seems to be working, but the other one seems to work fine for about 600 symbols and then suddenly all parameters, (coefficients, gain vectors, and other parameters defined in the new algorithm) suddenly all jump to infinity.

out of 1000 symbols say, the first 650, 640 or so are very good with errors only due to too much noise etc... but even if i suppress teh noise completely, then i still get this result and the erratic coefficients.